Traditional electric equipment cannot effectively detect the type of equipment and power start-stop information of impact load which has a great influence on the power grid. Therefore, a non-intrusive identification method of impact load based on siamese-architecture network is proposed. Firstly, the V-I trajectory characteristics and diagonal Gaussian harmonic characteristics of the waveform are extracted from the high-frequency sampling data at the inlet of the power edge equipment. On this basis, using the strong learning ability of convolutional neural network, a variety of prior information is preset to train the impact load characteristics of different equipment. In particular, a siamese-architecture network structure with shared network weight is designed to intelligently monitor and identify the occurrence of impact load and decompose its power by using different loss functions. The algorithm is deployed and tested based on ARM Cortex-M4 platform in a non-invasive way, and the identification ability of different identification algorithms for impact load was compared. The results show that the siamese-architecture network can be more accurate when high-power impact fluctuations occur in the power grid, the siamese-architecture network can more accurately identify the equipment category of the impact load and decompose its power, which effectively improves the identification effect of the impact load.